Page 16 - A U4SSC deliverable - Accelerating city transformation using frontier technologies

The deployment of autonomous vehicles which can observe their surrounding environments and
communicate with one another could improve road safety and, potentially, make traffic lights obsolete.
13
Despite its potential in the urban realm, the advancement of AI has also brought many uncertainties. The
replacement of jobs by AI is one of the most prominent concerns among citizens. A study conducted by
researchers at University of Oxford projected that up to 47 percent of total United States employment
is associated with occupations which could be automated in the coming decades. Automation could
also potentially displace 400-800 million jobs by 2030, which could require as many as 375 million
people to look for alternate employment.
14
Another major concern brought to light is that of data bias, which can also be reflected in AI algorithms.
Studies indicate that three distinct types of biases can be found in datasets: interaction bias – e.g.
people being misidentified in facial recognition due to lack of information; latent bias – e.g. people
being incorrectly identified based on historical data and stereotype; and selection bias – e.g. when a
dataset has overrepresented or underrepresented certain groups and swayed the selection process.
15
One of the most prominent examples of data bias in AI is the Correctional Offender Management
Profiling for Alternative Sanctions software, or COMPAS. COMPAS has been used by New York,
16
California, Wisconsin and other states in the US to assess the recidivism risk of criminals and support
related decision-making. Biased data has led to false predictions that African-American ethnicities are
twice as likely to reoffend as those of Caucasian origin.

“The advancement of AI has also brought many uncertainties. Jobs being
replaced by AI is among the most prominent concerns among citizens.”

Moreover, an emerging concern that is beginning to spill over and impede the deployment of AI is the
environmental efficiency aspect of their application. Using machine-learning methods to train an AI
model can be an energy-intensive process. Recent studies conducted at University of Massachusetts
Amherst suggest that training and searching a certain neural network architecture emits roughly 626
000 pounds of carbon dioxide, which is equivalent to five times the lifetime emission of the average
US car. 17

In order for AI to elicit desirable impacts, cities need to adopt an ethical framework to guide the
development of AI through an open and participatory process. Independent of the context, AI and
machine-learning models should be:

• lawful: respecting all applicable laws and regulations;
• ethical: respecting the ethical principles and values that are shared by all; and
• robust: taking the social environment into consideration.
18
Additionally, as cities continue to adopt AI solutions, successfully scaling up these solutions requires
cities to consider the environmental performance of AI and machine learning.

6 Accelerating city transformation using frontier technologies | A U4SSC deliverable